Economics ... Finance & Accounting ... Banking ... Fraud Detection ... Improper Payments
Prescriptive & Predictive Analytics ... Predictive Operations ... Forecasting ... with Excel ... Market Trading ... Sports Prediction ... Marketing ... Politics
Risk, Compliance and Regulation ... Ethics ... Privacy ... Law ... AI Governance ... AI Verification and Validation
Why We Shouldn’t Want Banks to Go All In on Artificial Intelligence ...Banks are spending lots on A.I., and they want to use it to manipulate you into spending more money. | Elena Botella - Slate

Banks were among the first large users of machine learning, and they now use AI in almost every part of the business. Four jobs matter most:

Credit and lending: deciding who gets a loan, at what price, and with what limit.
Fraud prevention: stopping stolen cards, account takeovers, scams, and fake identities, often in the seconds before a payment clears. See Fraud Detection.
Anti-money laundering (AML): finding and reporting money that criminals are trying to move or disguise, as the Bank Secrecy Act requires.
Customer service and operations: chat assistants, document processing, and internal tools for bank staff.

A fifth job is new: serving AI agents as customers. Shopping and payment agents can now search, decide, and pay on a person's behalf, so banks and card networks must work out how to recognize an agent, confirm what it is allowed to do, and decide who pays when it gets something wrong.

The main players are banks and credit unions, card networks (Visa, Mastercard), payment companies (Stripe, PayPal), AI developers (OpenAI, Google, Anthropic), and regulators. In the United States, the regulators include the Federal Reserve, the Office of the Comptroller of the Currency (OCC), the FDIC, the Consumer Financial Protection Bureau (CFPB), and Treasury's Financial Crimes Enforcement Network (FinCEN). Internationally, the Bank for International Settlements (BIS) Innovation Hub runs experiments on shared financial-crime analytics. For the broader business context, see Finance & Accounting.

The problem AI addresses is a mismatch of speed and scale. Money now moves in seconds, often irreversibly, across billions of transactions a day. Criminals use the same AI tools as banks, including deepfaked identity documents and voice clones. Meanwhile, the older approach of fixed rules (flag every cash transaction over a threshold, every transfer to a listed country) produces floods of alerts that are almost all false alarms.

The shift has come in three steps. Banks moved from rules to models that score risk from many signals at once. They are now moving from single-bank models to network analytics that look across institutions and payment systems, and from models that only flag activity to agents that assemble case files, draft reports, and, on the customer side, actually make payments.

In banking, AI now works on both sides of every transaction: it helps money move faster and decides who gets credit, and it helps stop money from reaching the wrong hands. The hardest problems are no longer only accuracy but trust: knowing who or what is acting, why a model decided, and who answers when it is wrong.

The Mechanism

Most bank AI systems follow the same loop, whether the goal is fraud, AML, or credit:

Collect signals. Transaction details, account history, device and location data, identity documents, credit bureau data, and, for AML, customer due diligence records.
Score the risk. Models estimate how likely an event is to be fraud, laundering, or default. Common techniques include gradient-boosted trees for tabular data, anomaly detection for unusual behavior, and graph or network analysis for links between accounts.
Decide. Scores combine with rules and thresholds. A payment may be approved, held, or blocked in real time; an AML alert goes to an analyst; a loan is approved, priced, or declined.
Investigate. Human analysts review alerts. Generative AI increasingly helps here by summarizing account activity, assembling case files, and drafting suspicious activity report (SAR) narratives, with a person accountable for the final decision.
Learn. Confirmed fraud, filed SARs, chargebacks, and loan outcomes become labels for retraining.
Govern. Models are validated, documented, and monitored under supervisory model risk guidance (the Federal Reserve and OCC's SR 11-7, issued in 2011), and credit models must produce explainable reasons for denials.

Agents add two new branches to this loop. On the bank side, investigation agents gather evidence across systems before a human reviews it. On the customer side, payment agents initiate transactions themselves, so the "collect signals" step must also verify the agent and the authority it was given.

Uncertainty remains at several points. Labels are sparse and noisy, because most AML alerts are never confirmed either way. Criminal behavior shifts as soon as a control works. And regulators are still deciding how much autonomy to allow in decisions that affect customers.

Constraints

Banking AI has no physics in the literal sense, but it does face hard quantitative constraints that shape every design choice.

Base rates make precision hard. Illicit transactions are rare. Suppose 1 in 10,000 transactions is illicit, and a model catches 99 percent of them while wrongly flagging just 1 percent of legitimate ones. In a million transactions, it catches 99 of 100 illicit ones but also flags 9,999 legitimate ones, so only about 1 percent of alerts are real. That is why industry estimates commonly put AML false-positive rates above 90 percent, and why small improvements in specificity save large amounts of analyst time.
Crime is a network, and each bank sees a fragment. Launderers spread money across many accounts and institutions. A single bank sees only its own slice. The BIS Innovation Hub's Project Aurora (2023) found that collaborative analysis across institutions detected laundering networks more effectively than siloed monitoring, and Project Hertha (2025) found that payment-system-level analytics helped banks identify 12 percent more illicit accounts, and 26 percent more for previously unseen behaviors. The constraint is privacy: sharing data across institutions requires legal authority (in the U.S., for example, voluntary sharing under Section 314(b) of the USA PATRIOT Act) or privacy-enhancing technologies.
Instant payments shrink the decision window. Real-time systems such as the Federal Reserve's FedNow service (launched July 2023) settle in seconds and are hard to reverse. Fraud models must decide in milliseconds, before the money is gone, which limits how much evidence they can gather.
The adversary adapts. Fraud is not a fixed distribution. When a control works, criminals change tactics, so model accuracy decays over time (concept drift). Generative AI speeds this up: deepfaked documents and voices are cheap to produce and improve quickly.
Errors cost different amounts. A missed fraud costs money; a false block costs a customer's trust and sometimes their access to their own funds. A wrongly denied loan can be unlawful discrimination. Banks tune thresholds to these asymmetric costs, and regulators require them to justify the trade-offs.
Explainability is a legal requirement, not a preference. Under the Equal Credit Opportunity Act and Regulation B, a lender that denies credit must give the applicant specific reasons. More complex models can be more accurate but harder to explain, so lenders use explanation methods (such as feature-attribution techniques) and sometimes accept simpler models.

Why This Matters

The stakes are large and growing:

The Federal Trade Commission reported that consumers lost $15.9 billion to fraud in 2025, up from $12 billion in 2024 and roughly 430 percent more than in 2020. Bank payments accounted for the largest losses by payment method, and investment scams for about half of all reported losses.
The FBI's Internet Crime Complaint Center reported about $20.8 billion in cybercrime losses for 2025, with crypto-related fraud above $11 billion.
The UN Office on Drugs and Crime has long estimated that 2 to 5 percent of global GDP is laundered each year. That figure is old and rough, but no better global estimate exists, which itself shows how hard the problem is to measure.

Better AI can cut losses and reduce wasted compliance work, and better credit models can extend lending to people with thin credit files. The same technology creates new risks: biased credit decisions, opaque denials, deepfake-driven account opening, and payment agents acting without clear authority. If network-level analytics and agent authorization standards work as intended, the industry could shift from each bank defending itself to shared defenses at the payment-system level. What remains uncertain is how to share data without eroding privacy, how to prove a model is fair, and who bears liability when an autonomous agent makes a mistake.

Recent Developments

Deepfake fraud warning (November 2024). FinCEN issued alert FIN-2024-Alert004 on November 13, 2024, warning that criminals use generative AI to create deepfake identity documents, photos, and videos to open accounts and move illicit funds. It listed red flags, including photos that look altered, a customer's apparent age not matching the date of birth, third-party webcam plug-ins during live verification, refusal of multifactor authentication, and reverse-image matches to galleries of AI-generated faces. It asked institutions to use the key term "FIN-2024-DEEPFAKEFRAUD" in related suspicious activity reports and recommended phishing-resistant multifactor authentication and live audio or video verification.

Treasury report on AI in financial services (December 2024). Treasury published a report based on public comments that described benefits (fraud detection, efficiency, credit access) and risks (data privacy, bias, explainability, and concentration among a few third-party AI providers). It is the most comprehensive federal statement on the topic so far.

Payment-system analytics for financial crime (June 2025). The BIS Innovation Hub's London Centre and the Bank of England published Project Hertha. Researchers built a synthetic dataset of 1.8 million bank accounts and 308 million transactions, generated by an AI model to mimic a national retail payment system, and tested whether analytics at the payment-system level could help banks spot financial crime. They found 12 percent more illicit accounts overall and 26 percent more for previously unseen behaviors. The result is a proof of concept on synthetic data, not a production deployment. The authors note that it depends on labeled training data, a working feedback loop, and explainable models, and that legal and regulatory questions were out of scope. It builds on Project Aurora (May 2023), which found that collaborative, cross-institution analysis outperformed siloed monitoring.

AML program rule reset (April 2026). On April 7, 2026, FinCEN proposed a rule to "fundamentally reform" AML/CFT program requirements. It focuses on effectiveness rather than paperwork, tells institutions to spend more resources on higher risks and fewer on lower ones, limits examiners' ability to substitute their own judgment for a bank's risk-based choices, and fully replaces FinCEN's July 2024 proposal. Comments were due June 9, 2026. The proposal does not address AI directly, but an effectiveness standard is likely to favor analytics that find more real crime with fewer false alerts. That is an interpretation, not a stated goal of the rule.

AI agents begin to pay (2025 to 2026). Card networks and technology companies released competing frameworks for payments made by AI agents in 2025: Mastercard Agent Pay and Visa Intelligent Commerce (April 2025), Coinbase's x402 protocol for crypto micropayments (May 2025), and Google's Agent Payments Protocol (AP2) and the OpenAI and Stripe Agentic Commerce Protocol (September 2025). Several moved into live use in early 2026; Santander and Mastercard reported completing Europe's first live end-to-end agent payment on production bank systems in March 2026. The Federal Reserve held an industry roundtable on agentic commerce on May 26, 2026, and Governor Christopher Waller laid out the open questions in a September 29, 2026 speech (see AI Agents and Payments).

Stablecoin law moves toward implementation (2025 to 2026). The GENIUS Act, enacted July 18, 2025, created the first federal framework for payment stablecoins. Regulators issued proposed rules through 2026: the OCC in March, Treasury in April and August, and the Federal Reserve in September. The OCC said it aimed to finalize its rule by November 2026. The law's main requirements take effect by January 18, 2027 at the latest. FinCEN and OFAC have proposed treating licensed stablecoin issuers as financial institutions with AML and sanctions obligations.

The Federal Reserve adopts AI internally and rebuilds data for agents (2026). In February 2026, Governor Waller described a shared AI platform for staff across all 12 Reserve Banks. In October 2026, the St. Louis Fed launched an MCP connector so AI agents can retrieve data from FRED directly (see The Federal Reserve Bank of St. Louis and AI Agents).

Money Laundering

Money laundering disguises the criminal origin of funds, traditionally in three stages: placement (getting cash into the financial system), layering (moving it through many transactions and accounts to obscure its source), and integration (bringing it back as apparently legitimate money). In the United States, the Bank Secrecy Act requires banks to know their customers, report cash transactions over $10,000 (currency transaction reports), and file suspicious activity reports when they see signs of illicit activity.

How AI Changes Anti-Money Laundering

Traditional transaction monitoring relies on fixed scenarios: a rule fires when, for example, several cash deposits fall just under the reporting threshold. Rules are easy to explain to examiners but generate many false positives and miss patterns nobody wrote a rule for. AI changes monitoring in four ways:

Entity resolution links records that refer to the same person or business across accounts, products, and spelling variations, so the system sees the whole customer.
Network analysis maps money flows between accounts to find rings of mule accounts, circular transfers, and shell-company structures that no single transaction reveals.
Risk scoring and alert triage rank alerts by likelihood of real suspicion, so analysts work the most important cases first and close obvious false alarms faster.
Investigation assistants use generative AI to summarize activity, pull together evidence from many systems, and draft SAR narratives. The analyst still reviews and signs off, because the bank is accountable for every filing.

The limits are real. Training labels are weak, because filing a SAR does not prove a crime and most outcomes are never reported back to the bank. Examiners expect models to be validated and explainable. And no single bank sees the whole network, which is why the BIS Aurora and Hertha experiments focused on cross-institution and payment-system analytics.

Fraud, Scams, and Deepfakes

Fraud and money laundering overlap: stolen and scam proceeds usually move through mule accounts, which are then used to launder them. Generative AI has lowered the cost of both. FinCEN's 2024 alert describes deepfaked identity documents used to open accounts, and voice and video impersonation used in scams that trick people into authorizing payments themselves. Authorized push payment scams are especially hard to stop, because the real customer approves the payment, so traditional authentication checks pass.

Banks are responding with layered defenses: liveness checks during identity verification, deepfake detection tools, device and behavioral signals (how a person types, holds a phone, or navigates an app), phishing-resistant multifactor authentication, and confirmation steps that slow down unusual first-time payments. None of these is decisive alone; the goal is to make fraud expensive enough that criminals look elsewhere.

Credit and Lending

Lenders use machine learning to estimate default risk, set prices and limits, and detect application fraud. Models can draw on more data than traditional credit scores, such as cash-flow data from bank accounts, which can help people with thin credit files qualify.

Two legal constraints shape how this works in the United States:

Adverse action notices. The Equal Credit Opportunity Act and Regulation B require lenders to tell declined applicants the specific principal reasons for the decision. The CFPB said in 2022 and 2023 that this applies even when a lender uses complex algorithms, and that generic checklist reasons are not enough. In 2025 the CFPB withdrew a large set of its earlier guidance documents, but the underlying Regulation B requirement remains law.
Fair lending. Models must not discriminate on prohibited bases such as race, sex, or age, including indirectly through variables that act as proxies. Lenders test models for disparate impact and search for less discriminatory alternatives that perform almost as well.

The open question is how far explanation methods can go. Feature-attribution techniques can say which inputs moved a score, but they do not always produce reasons a consumer can act on, and regulators and researchers still disagree about how reliable they are for complex models.

Blockchain, AI and Banking

Blockchain

Blockchain-based money is converging with mainstream banking in two ways. The first is stablecoins: tokens backed by safe assets and designed to hold a stable value, usually one dollar. The GENIUS Act (July 2025) requires U.S. payment stablecoin issuers to hold high-quality reserves, be licensed by federal or state regulators, and meet AML and sanctions obligations, with issuers of up to $10 billion outstanding allowed to choose state oversight. The second is tokenized deposits and tokenized settlement, which put existing bank money onto shared ledgers.

AI connects to both. Blockchain analytics firms use machine learning to trace funds across wallets and flag addresses linked to scams, sanctions, or ransomware, which matters because the FBI attributes more than half of reported 2025 cybercrime losses to crypto-related fraud. And some AI agent payment systems, such as Coinbase's x402, use stablecoins to make machine-to-machine micropayments that would be too small to send economically over card networks.

AI Agents and Payments

Agentic commerce means AI agents that shop and pay. In his September 29, 2026 speech at the Sibos conference, Federal Reserve Governor Christopher Waller distinguished two models:

Agent-assisted: the agent searches and recommends, but the person decides and pays.
Agent-delegated: the person gives the agent authority to buy within set limits.

He expects consumer purchases to come first, with agent-assisted use preceding delegated use, and sees business-to-business buying as a natural fit because it is recurring and rule-based, although higher values raise the cost of mistakes.

The industry frameworks take different approaches to the same problem of proving an agent is allowed to pay:

Framework	Backer	Approach
Mastercard Agent Pay	Mastercard	Registers agents and ties each payment to a verifiable record of the user's instructions, so out-of-scope purchases can be caught at the network level
Visa Intelligent Commerce	Visa	Extends Visa's tokenization and merchant network to agents; a related Trusted Agent Protocol handles agent identity
Agentic Commerce Protocol	OpenAI and Stripe	Uses single-use payment tokens limited to one merchant, one session, and an amount range
Agent Payments Protocol (AP2)	Google	Builds on agent-to-agent protocols and signed "mandates" that record what the user authorized
x402	Coinbase	Uses the web's "HTTP 402 Payment Required" response so an agent can pay for a resource with a stablecoin micropayment

Waller named the open issues: how to confirm an agent's authority rather than just the person's identity, who is liable when an agent buys the wrong thing, how to retrain fraud systems built around human behavior, how agents could raise cybersecurity risks, and how to support many small payments cheaply. He also asked whether standards will be interoperable or platform-specific, and whether agent systems will be open to many agents or locked to one. The Fed has hosted discussion but has not proposed rules.

The Federal Reserve Bank of St. Louis and AI Agents

The Federal Reserve Bank of St. Louis has become one of the most active parts of the Federal Reserve on artificial intelligence. Its work runs along two tracks: rebuilding FRED (Federal Reserve Economic Data), its economic database, so AI agents can use it reliably, and measuring how American workers actually use Generative AI.

FRED Is Being Rebuilt for AI Agents

On October 1, 2026, Federal Reserve Governor Christopher Waller, who previously led research at the St. Louis Fed, described the changes at FRED Con 2026, held at the St. Louis Fed.

Agents are now half of FRED's visitors. FRED traffic is growing about 150 percent a year, mostly from AI agents and bots. Roughly half of all visits now come from agents rather than people.
There is a FRED MCP Connector. It uses the Model Context Protocol (MCP), so users can point the AI agent of their choice at FRED to find and retrieve data. Staff describe this as "bring your own AI."
FRED staff are writing skills for agents. These help agents interpret economic data correctly. Waller said FRED was built to be read by humans and now has to be readable by machines as well.
Attribution is the main worry. Waller said agents have generally been accurate, but tests showed some agents miscrediting sources. FRED's value depends on being a trusted, neutral republisher of other organizations' data, so misattribution threatens that role.
More is planned. AI queries will be linked to FRASER, the St. Louis Fed's historical document archive, and to ALFRED, which keeps data as first reported so users can see later revisions. Requests for some Federal Reserve Board data will also be routed to FRED.

For scale, FRED holds more than 850,000 data series and had 18 million unique users in 2025. FRED itself has a long history: it began in 1961 as a typed weekly data digest, moved to a dial-up bulletin board in 1991, and launched on the web in 1995. The MCP connector is a working public example of the connector layer that agent environments use to reach outside data, and it lets a bank analyst's agent pull official economic series (interest rates, credit conditions, delinquency rates) without scraping web pages.

Measuring How Workers Use AI

The St. Louis Fed also publishes the Generative AI Adoption Tracker by economists Alexander Bick, Adam Blandin, and David Deming, added to FRED as data series in July 2026. It draws on a quarterly, nationally representative survey of adults aged 18 to 64. Between the third quarter of 2024 and the second quarter of 2026, the share of employed adults who used generative AI for work in the past week rose from 28.2 percent to 39.2 percent, and the share of work hours saved rose from 1.6 percent to 2.2 percent. The time savings are self-reported. A September 2026 St. Louis Fed study by the same team found AI use broad but shallow: common across occupations, but rarely used for most of any single job's tasks.

Federal Reserve-Wide Efforts

Internal AI. In February 2026, Governor Waller described a shared AI platform for staff at all 12 Reserve Banks. Every employee has access to approved AI tools, AI use is mandatory, AI skills are being written into performance goals, and hundreds of developers use coding assistants. He described agentic AI as "coming into view" rather than already deployed.
Agent payments. The Fed's work on agentic commerce is covered in AI Agents and Payments.

What the St. Louis Fed Is

The St. Louis Fed is one of the 12 regional Reserve Banks that, together with the Board of Governors in Washington, make up the Federal Reserve System. It serves the Eighth District (all of Arkansas and parts of Illinois, Indiana, Kentucky, Mississippi, Missouri, and Tennessee), with branches in Little Rock, Louisville, and Memphis, and it opened in November 1914.

Don't let the word "Federal" mislead you: the St. Louis Fed is not a federal agency in the way the Board of Governors is. It is a federally chartered, quasi-public corporation. Member banks in its district are required to buy its stock, but that stock cannot be sold or pledged, carries no ordinary ownership control, and pays dividends capped by law. Its board includes directors elected by member banks and directors appointed by the Board of Governors; since 2010, the banker directors may not vote on choosing the president, and the Board of Governors must approve the choice. The Reserve Bank operates under federal law, is supervised by the Board, and sends its profits to the U.S. Treasury after expenses and dividends. Courts have treated Reserve Banks as independent corporations rather than federal agencies for some purposes (for example, Lewis v. United States, 1982, on tort claims), even though they carry out public functions.

G. Edward Griffin’s 1994 book The Creature from Jekyll Island popularized the idea that bankers secretly created a private cartel. The documented history is narrower: in November 1910, Senator Nelson Aldrich and five financiers and officials met secretly at the Jekyll Island Club in Georgia and drafted a banker-led plan for a central bank. That plan failed in Congress. The Federal Reserve Act of 1913 kept its regional structure but added a presidentially appointed board and more public control, so the claim that the Fed as enacted is a private cartel is not supported, although critics still debate how much influence bankers have on the regional boards. The Meeting at Jekyll Island, Federal Reserve History

Additional Viewing and References

Additional Viewing

The embedded videos on this page appear in the sections above. These recent talks, available on their official pages, cover the newest developments discussed here:

Payments in the Age of AI Agents | Governor Christopher Waller, Federal Reserve Board Waller's September 29, 2026 Sibos speech on agent-assisted and agent-delegated payments, authentication, liability, fraud, and standards. It is the clearest regulator's map of the open questions in agentic commerce.
The Data Version of Godzilla versus Kong: FRED Takes on AI | Governor Christopher Waller, Federal Reserve Board Waller's October 1, 2026 FRED Con remarks on AI agent traffic, the FRED MCP Connector, and attribution concerns.
Operationalizing AI at the Federal Reserve | Governor Christopher Waller, via BIS Waller's February 24, 2026 remarks on how the Federal Reserve deploys AI internally, with its guardrails.

References

Financial crime and fraud

FinCEN Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions | FinCEN November 13, 2024 alert describing deepfake identity fraud, red flags, and the SAR key term.
FinCEN Proposes Rule to Fundamentally Reform Financial Institution Programs Designed to Fight Illicit Finance | FinCEN April 7, 2026 proposal shifting AML/CFT programs toward effectiveness and risk-based resource allocation.
Project Hertha: Identifying Financial Crime Patterns in Real-Time Retail Payment Systems | BIS Innovation Hub and Bank of England June 2025 proof of concept: 12 percent more illicit accounts identified, 26 percent more for unseen behaviors, on synthetic data.
Project Aurora: The Power of Data, Technology and Collaboration to Combat Money Laundering | BIS Innovation Hub May 2023 study finding that collaborative, cross-institution analytics outperform siloed monitoring.
Consumer Fraud Losses Quintupled Since 2020, FTC Says | PYMNTS Coverage of the FTC's March 2026 testimony reporting $15.9 billion in 2025 consumer fraud losses.
Cybercrime Losses Break the $20 Billion Mark | Help Net Security Coverage of the FBI IC3 2025 Internet Crime Report.

Payments, agents, and stablecoins

Payments in the Age of AI Agents | Governor Christopher Waller, Federal Reserve Board September 29, 2026 speech on agentic commerce.
Mastercard Agent Pay vs Stripe ACP vs Visa Intelligent Commerce vs x402: A 2026 Update | ATXP A vendor's comparison of agent payment frameworks and their 2026 status; useful but not independent.
Implementing the GENIUS Act for Stablecoin Issuers Under OCC Jurisdiction | OCC, Federal Register March 2026 proposed rule for federally supervised stablecoin issuers.
GENIUS Act Stablecoin Rules: Status up to 2027 | CryptoTicker Timeline of GENIUS Act rulemaking through October 2026.

The Federal Reserve and AI

The Data Version of Godzilla versus Kong: FRED Takes on AI | Governor Christopher Waller October 1, 2026 remarks on FRED and AI agents.
St. Louis Fed Adapts FRED Database for AI Agents | PYMNTS Coverage of the FRED MCP Connector announcement.
Fed Adapting Data Tools for AI Bots | American Banker Coverage of FRED's changes for AI agents.
Does Generative AI Save Time at Work? | FRED Blog, St. Louis Fed August 2026 summary of the Generative AI Adoption Tracker.
What Work Does Generative AI Do? | Bick, Blandin, Deming, and Schumacher, St. Louis Fed September 2026 task-level study of AI use across occupations.
Operationalizing AI at the Federal Reserve | Governor Christopher Waller, via BIS February 2026 remarks on the Fed's internal AI platform.
The Meeting at Jekyll Island | Federal Reserve History The Federal Reserve's own account of the 1910 meeting and the Aldrich Plan.

Banking